# A Beginner’s Guide to Large Models: The Essentials Willdom Lee · 2025-03-08 [HTML](https://willdomlee.com/en/articles/llm-basics) · en > When answers are within easy reach, the wisdom to ask good questions and the courage to act matter even more. > > Reducing a model's hallucinations can also reduce its creativity. Innovation needs ideas to emerge, but not every idea that emerges is meaningful. ## What everyone should know about large models News and technical breakthroughs in large models arrive so quickly that it is hard to keep up. Yet their underlying framework and core principles have not fundamentally changed. As models become more capable, creating real business value still requires high-quality knowledge, ongoing data governance, new ways of thinking, and the ability to guide the models well. In a time of change, we need to distinguish what changes from what endures: rivers keep flowing, while mountains stand still. The water moves; the mountain does not. Today, I want to discuss some of the things that are not moving quite so quickly. I have picked out concepts I think everyone should understand: **large models, parameters, pretraining, fine-tuning, generalization, prompts, inference, mixture of experts (MoE), and hallucinations**. Let's explain them as simply and plainly as possible. ![](https://willdomlee.com/assets/a/llm-basics/01.webp) Key concepts in understanding large models **Large models:** The most common are large language models, or LLMs. By analyzing patterns in enormous amounts of text—on the scale of human civilization—they develop an ability to understand the world through language. Think of a machine with billions of adjustable knobs. There are other kinds of large models too, such as image and video generators, which fall under multimodal models. **Parameters:** These are the knobs inside the machine, numbering in the billions and now even trillions. Training continually adjusts them, encoding what the model learns from data by finding patterns rather than simply memorizing. More parameters mean greater capability and greater resource use. Bigger is not always better; the size needs to fit the task. The diagram below is one I found that is both intuitive and reasonably rigorous. The input—the question—is at the beginning; the output—the answer—is at the end. The lines in between represent a simplified neural network. ![](https://willdomlee.com/assets/a/llm-basics/02.webp) A schematic of a large model **Pretraining:** Initially, those countless knobs are set randomly, and the model knows nothing. Pretraining is like sending it to school with a large collection of good textbooks until it acquires general knowledge roughly comparable to an undergraduate education. It consumes vast resources and puts heavy demands on computing and networking. In Chinese tech circles, people jokingly call it alchemy. ![](https://willdomlee.com/assets/a/llm-basics/03.webp) Training large models is often nicknamed alchemy. **Fine-tuning:** Give the model specialist textbooks and keep training it until it becomes an expert in that field, rather like progressing to a master's degree or PhD. ![](https://willdomlee.com/assets/a/llm-basics/04.webp) The model's schoolbooks; on the right is an actual pretraining corpus. **Generalization:** The ability to apply what it has learned to unfamiliar questions and combine ideas to create knowledge. Training discovers patterns across vast amounts of knowledge rather than merely storing it. **Prompts:** Instructions that tell a model what you want it to do—for example, “Explain quantum mechanics in terms a primary school student can understand.” Writing prompts is itself a skill in guiding a model. **Inference:** After receiving a prompt, the model works on the task and produces a result. This uses far fewer GPU resources than pretraining. For networking, 25G and 100G RoCE Ethernet switches can meet most requirements. ![](https://willdomlee.com/assets/a/llm-basics/05.webp) Prompts and inference **Mixture of experts (MoE):** Think of departments collaborating inside a model. Different tasks automatically use different submodels—A for translation, B for coding—rather than using every knob all the time. This improves efficiency. ![](https://willdomlee.com/assets/a/llm-basics/06.webp) To calculate 1 + 1, only one expert needs to be activated! **Hallucinations:** A model confidently talking nonsense: producing something that fits its internal patterns but not reality. Yet reducing hallucinations can also reduce creativity. Innovation needs ideas to emerge, but not every idea that emerges is meaningful. ![](https://willdomlee.com/assets/a/llm-basics/07.webp) An actual GPT hallucination at the time, since fixed Just as humans need language to think, large models build a framework for understanding the world through vast numbers of adjustable parameters. We first train them by having them read widely, then fine-tune them with specialist practice. They gain the ability to generalize and apply what they have learned to new problems. We direct them with prompts, and they perform inference to answer. Sometimes they hallucinate, which is where human supervision matters—or perhaps they are producing a better answer that humans have not yet recognized. The result of training—the final elixir from this alchemy—is a set of patterns distilled from knowledge. Each knob's position becomes a weight. It is not the knowledge itself, and humans cannot even read what is inside that elixir. The model uses these patterns to fit the world and compress the knowledge of human civilization. It is a remarkable thing. ## Why are so many industries deploying private models, especially DeepSeek? - DeepSeek, along with recent open models from Alibaba such as Qwen and QwQ, is genuinely smart—a source of pride for Chinese technology. - Low costs for training and inference, open-source availability, and the security of keeping private data local. - Real industry needs and opportunities to create value in healthcare, finance, government, business, and education. AI can support operations, establish an early position, build datasets, accumulate knowledge, and develop the ability to guide AI. - It broke into the mainstream over Spring Festival, creating a constructive wave of adoption and competitive pressure—much as every organization once felt it needed a website. - Policy and capital are helping drive adoption, alongside an opportunity for strategic change. - **At heart, science and technology are the primary productive forces.** ![](https://willdomlee.com/assets/a/llm-basics/08.webp) More next time! If you would like to keep reading, here are a few earlier posts: [Is Training AI Really That Hard? I Tried a 26M-Parameter Model](https://willdomlee.com/en/articles/training-a-tiny-model.html) [A Few Thoughts on ChatGPT](https://willdomlee.com/en/articles/on-chatgpt.html) --- Originally published in Chinese on WeChat, 南山独唱团. [Read the original](https://mp.weixin.qq.com/s/2Bg2t8LZpY64cnZ6ckZujw).